High-speed digital communication represents an increasingly important technology. An increasing number of analog communication links are replaced with high-speed serial links offering reliability, link quality (e.g. low Bit Error Rate (BER)), simplicity of use and low cost to end users.
For example, modem computer displays and High-Definition Television (HDTV) displays often employ high-speed serial links based on industry standards such as Digital Visual Interface (DVI) and High Definition Multimedia Interface (HDMI). The aggregate data rate for such links is generally in the Gigabits/second range. For example, the DVI link baud rate approaches 5 Gigabaud, with many vendors offering products exceeding this limit by non-trivial margins.
Another example is a serial Advanced Technology Attachment (ATA) interface for connecting storage devices in computer systems. This application also requires data rates of about 2 Gigabits/second and above.
Still other examples are high-speed Local Area Network (LAN) applications such as Gigabit Ethernet (such as 1000Base-T) as well as emerging 10-Gigabit Ethernet (10GBASE-T) standards.
The links underlying these and similar examples have many common requirements. They require high-speed operation, an ability to tolerate substantial deterministic and/or non-deterministic data jitter, an ability to tolerate substantial frequency offset and/or frequency modulation (FM), good effective BER, an ability to operate with small horizontal eye openings (for example a 0.4 unit interval (UI) eye width specification is not uncommon), high tracking bandwidth, fast phase acquisition and good tracking range, to name a few.
One popular implementation of a receiver for such a high-speed serial link is based on receiver data over-sampling combined with so called ‘Phase picking’ to select ‘good’ sample streams. Common DVI link implementations use an over-sampling factor of L=3, and data sampling is often done without any phase tracking. Once the over-sampled data stream is available, the phase picker retains one sample out of L samples by using some type of phase tracking approach (such as transition phase averaging) in order to select the optimal sample.
While such a technique is relatively simple and inexpensive to implement, it suffers from a variety of problems. For instance, some line codes employed today have very wide frequency spectrums (i.e. poor run length (d,k)-constraints). This problem is especially acute for DVI/HDMI links which can have runs from 1 to 14 symbols. For reasons of stability and robustness, phase pickers have to operate with relatively narrow tracking bandwidths. The downside of such an approach is an inability to track deterministic jitter present in the over-sampled data. Narrow tracking bandwidth also places limitations in settling speeds (i.e. phase acquisition times), causes poor transient performance and sometimes narrows the capture range.
One fundamental problem of the phase picking approach is its inability to use all available over-sampled data in the detection process. Common phase picker implementations simply discard L-1 samples for each symbol, retaining just one out of L received samples. Hence, the over-sampled data is used only for phase tracking and not for data detection. Thus, the phase picker discards many data samples which carry useful information about the symbol which is being decoded. In common DVI/HDMI implementations, about ⅔ of the information available to the receiver is simply discarded.
If the data rate is moderate and the eye opening is adequate, such as when the link employs a short high-quality Shielded Twisted Pair (STP) cable, the loss caused by such a phase picking approach is tolerable and makes for an inexpensive receiver implementation. However, once the data rate is increased and/or a longer or lower quality cable is used, such as an Unshielded Twisted Pair (UTP) cable, the available eye width becomes smaller and receiver performance begins to degrade rapidly.
For example, it can be shown that such a traditional phase-averaging phase picker requires an eye opening that is greater than 2 sampling intervals in order to reliably detect data symbols. In case of a common implementation using L=3 samples, this implies that the eye width should be at least 0.7UI if data errors due to phase picking are to be avoided. Such a limitation is quite acceptable at low speeds and/or short transmission distances, but becomes an expensive luxury at higher data rates and often requires good equalization and pre-emphasis techniques which are relatively expensive.
In addition, many traditional phase picking approaches are difficult to implement at high speeds and/or high over-sampling factors because they generally do not provide an easy way for parallel-processing of the samples. From a receiver implementation standpoint, it is often desirable to have the receiver complexity depend linearly on the over-sampling factor, thus providing for narrower eye specification. Many existing phase picking methods do not exhibit such a linear increase in complexity and instead grow with the square of the over-sampling factor or even more.
An alternative class of receivers well known in the art is represented by so called ‘Majority Voting’ over-sampled receivers. This technique was widely used in the past for low speed links and offers some linear and/or non-linear filtering of the data, often improving the robustness and BER of the link, provided a good eye opening is available, such as an eye opening of 0.9 UI or wider. Such receivers are also inexpensive to implement, which is an important reason for their popularity in the past. However, once the eye opening begins to shrink, as is common is today's high-speed serial links, the performance of such a ‘Majority Voting’ receiver quickly becomes unsatisfactory. Furthermore, in general such a receiver has difficulty handling data streams with high frequency offset and frequency modulation, as well as data streams with high non-deterministic jitter.
Compared to traditional ‘analog’ clock and data recovery (also known as CDR) techniques, data detections and/or phase tracking in the digital domain often have the advantages of portability between different fabrication processes, simple design and low cost. In addition, many non-linear detection and/or filtering methods that might be unavailable for analog implementation are often inexpensive to implement in the digital domain. Furthermore, even though conventional analog CDR often provides good phase tracking performance, it seldom contributes to improved data detection per se; most analog CDRs have a simple sampler for data detection that takes just one sample per symbol. Thus, if substantial eye closure due to high deterministic jitter is present, analog CDR performance becomes unsatisfactory.
In general, the best receiver operation performance may be achieved if every symbol is detected (i.e. sampled) at the point of maximum eye opening height. However, such a sampling point is difficult, and often impossible, to find. In order to determine an ‘optimal’ sampling point, a receiver usually consults data transitions for guidance. However, such data transitions generally do not occur right at symbol boundaries. Their position is affected by deterministic jitter (caused by inter-symbol interference (ISI) determined by the history of preceding symbols), non-deterministic jitter introduced by noise in various components of the transmission system (such as clock generator phase noise, imprecise slicing of signal by sampler resulting in amplitude and time offsets, additive and multiplicative channel noise, etc.), crosstalk-induced jitter and frequency offset and frequency modulation often present in such a link. In addition, the link often suffers from nonlinear distortions that further complicate reception.
Because the location of transitions may not provide reliable information regarding an optimal sampling point for the current symbol, the receiver may attempt to filter the sampling phase information derived from the observation of the transitions in order to arrive at some acceptable sampling phase. For instance, many existing implementations apply either a linear filter with low-pass transfer characteristics or some nonlinear filter with noise suppression properties (e.g. a median filter). However, the sampling phase position obtained with such a method is often non-optimal in the symbol-by-symbol sense. This is because it is difficult to separate deterministic components (such as ISI and crosstalk) from pure non-deterministic phase noise and frequency offset. While non-deterministic phase noise should be filtered out (i.e. suppressed), the deterministic part as well as the frequency offset should be tracked in order to provide good receiver performance. A filtering approach is often unable to make such a distinction and thus, overall detection performance suffers.
There are other well-known methods that attempt to improve the performance, such as sequence detection method based on the Viterbi Algorithm (VA) and its modifications, as well as feedback detection schemes and Decision Feedback Equalization (DFE). These approaches work well and allow to account for ISI and other deterministic components, thus improving overall performance and simplifying the sampling phase tracking task (since it now only needs to handle non-deterministic parts as well as frequency offsets). However, sequence-detection based methods often are complex and expensive to implement, generally requiring multi-bit sampling (i.e. analog-to-digital conversion (ADC)) and extensive computations on a symbol-by-symbol basis. Once the data rate starts to approach the Gigabit/second range, sequence detectors quickly become impractical. Even at low rates, VD (Viterbi Decoding) and DFE based receivers are prohibitively expensive to implement.
Therefore, what is needed is a simple yet efficient way to build a data receiver with over-sampled data reception which doesn't suffer from the problems which plague traditional phase pickers, ‘Majority Voting’ and analog CDR receivers.